Modeling Annotation Delay In Continual Learning

Botos Csaba, Wenxuan Zhang, Matthias Müller, Ser-Nam Lim, Mohamed Elhoseiny, Philip Torr, Adel Bibi

Primary Area: transfer learning, meta learning, and lifelong learning

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Keywords: continual learning, unsupervised learning, label delay

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.

Abstract:

Online continual learning, the process of training models on streaming data, has gained increasing attention in machine learning. However, a critical aspect often overlooked is the label delay, where new data arrives before its corresponding labels due to slow and costly annotation processes. We introduce a novel continual learning setup that explicitly accounts for label delay, with models trained over time steps with labels shifted from the data streams by some factor. In each step, the model is exposed to both unlabeled data from the current time $t$ step and labels from time step $t-d$. We show that this is a challenging problem and increasing the per step computational budget can not help resolve the problem. Moreover, we show that Self-Supervised learning and Test-Time Adaptation approaches perform even poorly compared to a naive approach that ignores the unlabaled data training on the older but annotated stream. We introduce a simple, efficient baseline using importance sampling to align the unlabeled and labeled data distributions, bridging the accuracy gap caused by label delay without significantly increasing computational complexity, by rehearsing from memory labeled samples that are most similar to the new unlabeled samples. While on CLOC our method performs similarly to SSL and TTA, on CGLM, our method not only closes the accuracy gap, but outperforms the non-delayed counterpart by $+25\%$ up to $+33\%$, while being computationally friendly.
We conduct various ablations and sensitivity experiments demonstrating the effectiveness of our approach.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 1418